基于公理模糊集与粒计算的人脸语义提取方法

PDF下载HTML

Title:

A facial semantic extraction method based on axiomatic fuzzy sets and granular computing

作者:

任艳¹, 张茜²

1. 沈阳航空航天大学 人工智能学院，辽宁 沈阳 110136;
2. 沈阳航空航天大学 自动化学院，辽宁 沈阳 110136

Author(s):

REN Yan¹, ZHANG Qian²

1. College of Artificial Intelligence, Shenyang Aerospace University, Shenyang 110136, China;
2. College of Automation, Shenyang Aerospace University, Shenyang 110136, China

关键词:

语义提取; AFS理论; 信息粒; 聚类; 人脸检索; 关键点检测; 隶属函数; 公理模糊集

Keywords:

semantic extraction; AFS theory; information granules; clustering; face retrieval; critical point detection; membership function; axiomatic fuzzy sets

分类号:

TP391.4

DOI:

10.11992/tis.202108023

文献标志码:

2022-05-20

摘要:

在人脸检索和验证领域，人类更倾向于通过描述对象特征的“语义”或“概念”来对人脸进行相似性判别，而传统的图像检索已无法满足这一需求。因此，本文提出了一种基于公理模糊集(axiomatic fuzzy sets, AFS)与信息粒的人脸语义提取方法(IAFSGD)。首先，对人脸图像进行校正并检测人脸关键点，进而基于关键点提取人脸特征；然后，对人脸面部特征样本进行聚类，构建类中心，在AFS框架下求取每类的信息粒，并通过得到的信息粒对人脸图像再次进行分类，从而得到最终的聚类结果和具有可解释性的面部语义描述；最后，将本文提出的算法在Multi-PIE、AR、FEI人脸数据库进行实验验证。实验结果表明，与FCM(fuzzy c-means)、CAN(clustering with adaptive neighbors)、FCMGD、AFSGD、KTM(K-means tree)算法相比，本文提出的语义提取方法可以获得与人类感知更为接近的聚类结果，且结果具备很好的可解释性。

Abstract:

People prefer to distinguish the similarity of faces in the field of face retrieval and verification by describing the “semantics” or “concept” of object features, which cannot be satisfied by traditional image retrieval technology. Therefore, this paper presents a facial semantic extraction algorithm (IAFSGD) that is based on axiomatic fuzzy sets (AFS) and information granules. First, the face images are corrected to detect the critical points where facial features are extracted; then, the samples of facial features are clustered to construct the class center; the information granules of each class are obtained under the AFS framework; and then the facial images are reclassified through the information granules, to obtain the final clustering results and interpretable facial semantic description; finally, the efficacy of this algorithm is demonstrated on such facial datasets as Multi-PIE, AR, and FEI. Experimental results show that compared with FCM, CAN, FCMGD, AFSGD, and KTM, the semantic extraction method proposed in this paper can obtain clustering results closer to human perception and that the results are of good interpretability.